STAIN: Spatial-Temporal Adversarial Network for Multivariate Time Series Imputation
Guangyu Liu, Pengfei Shen, Zhanguo Ma, Min Hyun Han, Nannan Lu · 2024
In the collection of time series data, missing values unavoidably occur due to equipment failures and human errors, which challenge the subsequent analysis and modeling. Thus, imputation becomes a common solution. Previous studies primarily focus on extracting temporal relationship for data generation, while neglect the relationship between variables. Such oversight can introduce biases into the imputed data and affect the reliability of subsequent time series analysis. In this study, we propose a Spatial-Temporal Adversarial Imputation Network (STAIN) for time series. STAIN leverages the generative adversarial network to acheive the imputation. However, it replaces the random noise with the mean imputation as pre-imputation for more accurate approximation of the ground truth. Then, a bilevel attention mechanism is proposed to capture the temporal and spatial dependencies of missing values with others in multivariate time series. Meanwhile, the discriminator with hint information is introduced as an adversary to encourage the sequence generator to generate more realistic outputs. Extensive experimental results demonstrate that STAIN can effectively improve the imputation accuracy and achieve the state-of-the-art results on three real-world datasets.